MiMo-V2.6 could keep calling tools without making progress, filling its context and leaving users waiting. In a September 27 postmortem, Xiaomi describes how that behavior developed during training and how it changed the model to stop it.
The team separates genuine parallel work from repetition. Reading several different files can be useful; repeatedly issuing the same call without new information is the failure under investigation. Its reported metric counts exact duplicate calls inside a single turn, so it misses some near-duplicates and loops across turns.
A training penalty originally activated only when a turn exceeded 32 calls. Xiaomi found that excessive activity below that threshold could grow before the penalty applied. Tightening the cap reduced call volume, but the fix was expensive to train and did not eliminate repetition across its evaluation sets.
Xiaomi instead trained a specialized model to avoid repetition and incorporated that behavior into the main checkpoints. It reports substantial reductions across the tested tools and context lengths while broader benchmark performance held steady.
The revised API models have been served under the existing names since September 25 at 06:00 UTC+8. Downloadable checkpoints carry an MOPD suffix. September 27 is the diagnostic report’s date, not the start of the API rollout.